A multi-objective radiomics model for the prediction of locoregional recurrence in head and neck squamous cell cancer

Purpose: Locoregional recurrence (LRR) is the predominant pattern of relapse after nonsurgical treatment of head and neck squamous cell cancer (HNSCC). Therefore, accurately identifying patients with HNSCC who are at high risk for LRR is important for optimizing personalized treatment plans. In this work, we developed a multi-classifier, multi-objective, and multi-modality (mCOM) radiomics-based outcome prediction model for HNSCC LRR. Methods: In mCOM, we considered sensitivity and specificity simultaneously as the objectives to guide the model optimization. We used multiple classifiers, comprising support vector machine (SVM), discriminant analysis (DA), and logistic regression (LR), to build the model. We used features from multiple modalities as model inputs, comprising clinical parameters and radiomics feature extracted from X-ray computed tomography (CT) images and positron emission tomography (PET) images. We proposed a multi-task multi-objective immune algorithm (mTO) to train the mCOM model and used an evidential reasoning (ER)-based method to fuse the output probabilities from different classifiers and modalities in mCOM. We evaluated the effectiveness of the developed method using a retrospective public pretreatment HNSCC dataset downloaded from The Cancer Imaging Archive (TCIA). The input for our model included radiomics features extracted from pretreatment PET and CT using an open source radiomics software and clinical characteristics such as sex, age, stage, primary disease site, human papillomavirus (HPV) status, and treatment paradigm. In our experiment, 190 patients from two institutions were used for model training while the remaining 87 patients from the other two institutions were used for testing. Results: When we built the predictive model using features from single modality, the multi-classifier (MC) models achieved better performance over the models built with the three base-classifiers individually. When we built the model using features from multiple modalities, the proposed method achieved area under the receiver operating characteristic curve (AUC) values of 0.76 for the radiomics-only model, and 0.77 for the model built with radiomics and clinical features, which is significantly higher than the AUCs of models built with single-modality features. The statistical analysis was performed using MATLAB software. Conclusions: Comparisons with other methods demonstrated the efficiency of the mTO algorithm and the superior performance of the proposed mCOM model for predicting HNSCC LRR.